基于典型相关分析和距离度量学习的零样本学习

摘要

零样本学习是一类特殊的图像分类问题,是指测试数据的类别在训练数据中没有出现的情况.为了更好地描述语义特征空间中图像特征和语义特征的距离关系,本文将距离度量学习引入零样本学习任务.具体而言,首先利用典型相关分析将样本的图像特征和相应类别的语义特征映射至公共特征空间;然后,利用距离度量学习衡量图像特征和语义特征之间的距离;最后,使用最近邻分类器进行分类.通过在流行的AwA和CUB数据集中的实验,证明了所提方法的有效性和鲁棒性.%Zero-shot learning is a special case of image classification,whose test classes are absent in training sam-ples.To better measure the distance between visual features and semantic features in the semantic embedding space, a distance metric learning based zero-shot learning method is proposed.Specifically,visual features and semantic features were first projected into a common semantic embedding space by use of canonical correlation analysis,then a distance metric learning method was employed to measure the distance between them.Finally,a nearest neighbor classifier was utilized to perform the classification.Experimental results on the popular AwA and CUB datasets dem-onstrate that the proposed approach is effective and robust.